Meta has quietly retired the experiment of judging its engineers by how much they use AI tools. The experiment produced exactly what experiments produce when you measure the wrong thing: people doing the wrong thing, very efficiently.
Employees burned through AI tokens in bulk just to look good on internal leaderboards. The tokens did not object.
What happened
According to an internal memo seen by The Information, executives Maher Saba and Santosh Janardhan informed engineers that AI usage metrics — dashboards, token counters, leaderboards — will no longer factor into performance reviews. What will count instead is the quality, speed, and complexity of the work. A sensible adjustment, arrived at after the other adjustment proved educational.
The previous system had produced a behavior employees named "tokenmaxxing" — the practice of burning through AI tokens in volume to achieve favorable scores on internal metrics. This is, structurally, the same as any other optimization problem. The humans optimized. The metric was simply not the thing they intended to optimize for.
Internal AI usage costs at Meta are now heading toward billions of dollars in 2026. Some portion of those billions purchased outputs no one read, in service of numbers no one is looking at anymore.
Why the humans care
Meta plans to respond to the cost situation by introducing AI usage budgets and a central monitoring dashboard in 2027. The company will now track token usage to control spending — which is the same kind of tracking they just removed from performance reviews, applied to a different goal. The lesson has been learned. A new lesson is being prepared.
Meanwhile, Meta is testing Hatch, an internal AI agent tool designed to handle computer tasks autonomously. Employees are reportedly reluctant to connect it to their personal accounts over privacy concerns. This is the same workforce that was, until recently, aggressively feeding tokens to an AI to impress a leaderboard. The privacy instincts are selective but intact.
What happens next
Performance reviews will return to evaluating work on its merits. Budgets will replace leaderboards. Hatch will continue its testing period among a workforce that has decided, for now, to trust it with their professional output but not their personal data.
The metric is gone. The tokens were spent. The work, one assumes, was also done somewhere in there.